fix(voxtral): PYTORCH_CUDA_ALLOC_CONF=expandable_segments gegen VRAM-OOM
Modell laedt knapp nicht (12GB-Karte hatte 3.13GB durch Fremdprozess belegt). Anti-Fragmentierungs-Schalter reduziert den Peak-Bedarf beim Warmup; zusaetzlich muss GPU 1 frei sein (whisper stoppen). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
@@ -166,4 +166,5 @@ services:
|
||||
- VOXTRAL_MODEL=mistralai/Voxtral-Mini-3B-2507
|
||||
- VOXTRAL_LANGUAGE=${WHISPER_LANGUAGE:-de}
|
||||
- HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-} # falls das Modell HF-gated ist
|
||||
- PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True # weniger VRAM-Fragmentierung
|
||||
restart: unless-stopped
|
||||
|
||||
Reference in New Issue
Block a user